Rollout of 2026-10-04: ChemAI and TinkerCliffs#

Paul’s explicit ask (2026-10-04): update ChemAI’s and TinkerCliffs’ SEAMM installations to the current releases (phase 4, the PBS chain, seamm_exec 2026.10.4, read_structure_step 2026.10.4). Done the same day.

Versions#

Package

ChemAI before

ChemAI now

TinkerCliffs before

TinkerCliffs now

seamm

2026.10.2

2026.10.3

2026.10.2.2

2026.10.3

molsystem

2026.9.25

2026.10.3

2026.9.25

2026.10.3

seamm-exec

2026.9.27

2026.10.4

2026.10.2.1

2026.10.4

seamm-scheduler

(none)

2026.10.3

2026.10.2

2026.10.3

seamm-jobserver

2026.9.27

2026.10.3

2026.10.2

2026.10.3

seamm-slurm

2026.8.13

2026.10.2

2026.10.2

2026.10.2

seamm-webui (own venv)

2026.10.1.1

2026.10.3

(no web UI)

–

table-step

2026.9.30

2026.10.3

2026.9.30

2026.10.3

loop-step

2026.9.18

2026.10.3

2026.9.18

2026.10.3

geometry-analysis-step

2026.3.1

2026.10.3.1

2026.3.1

2026.10.3.1

orca-step

2026.10.1

2026.10.3.2

2026.10.1

2026.10.3.2

mopac-step

2026.10.1

2026.10.3.2

2026.10.1

2026.10.3.2

read-structure-step

2026.9.18.1

2026.10.4

2026.9.18.1

2026.10.4

vasp-step

2026.9.29

2026.10.3.1

2026.9.29

2026.10.3.1

lammps-step

2026.10.1

2026.10.3

–

–

xnn-step

2026.9.28

2026.9.28

2026.9.28

2026.10.2

seamm-manager (tool)

2026.10.1.1

2026.10.2.3

2026.10.2.3

2026.10.2.3

pip check is clean on both. properties-step is not installed on either (update only upgrades what is installed). mbe-step 2026.10.4 and seamm-mbe 2026.10.3.1 were added to ChemAI afterwards (below); TinkerCliffs has neither.

ChemAI#

  1. No jobs: nothing in SLURM for seamm, nothing submitted or running in the datastore (the 112 old “started” rows are stale). JobServer and web UI stopped.

  2. seamm-manager 2026.10.2.3 (uv tool install --force --python 3.12).

  3. environment migrate: the old venv is venvs/2026-10-04T06-14-09 (the rollback). It reported starting the JobServer; it was not running (seamm_manager#26).

  4. Not update --all: xnn-step’s installer would have applied seamm-xnn.yml with conda env update to seamm-lammps, which xnn.ini names and which has no xnn applied-marker – the LAMMPS/MLFF environment the rollout must not touch. Paul chose to update everything except xnn-step: update --latest with every installed package but xnn-step, plus seamm-webui. New venv venvs/2026-10-04T06-18-41 (19 changes). The other code installers found their environment files unchanged since 2026-09-28 (markers under 7 days old) and did nothing: every code environment’s marker and conda-meta are unchanged; seamm-lammps’ history was last written 2026-09-19. The pyxtal-step installer failed (import pkg_resources), a problem of that plug-in, not of the update.

  5. JobServer 2026.10.3 started (services start jobserver), web UI running (55155) with all 16 queues; the legacy dashboard left stopped.

TinkerCliffs#

/projects/seamm/SEAMM was already versioned with seamm-manager 2026.10.2.3. None of the jobs running there used it (the MBE session’s bundles use its private venv; the qzDimer2 array runs ORCA directly), so update --latest --all built venvs/2026-10-04T06-12-15 and switched; the previous version venvs/2026-10-03T06-33-37 is the rollback.

Damage: --all included xnn-step, whose installer applied seamm-xnn.yml to the env xnn.ini names, seamm-lammps-xnndev (the ARC MLFF env). At 06:14-06:15 that upgraded openssl 3.6.4 -> 3.6.5 (conda) and replaced xnns 0.5.0 from git a8958df6 with xnns 0.6.0 from PyPI (pip, xnns>=0.4.0), and wrote the seamm-xnn-step/seamm-lammps-step markers. The development session had installed 0.6.0 at 05:58, but into the separate clone seamm-lammps-xnn060; its pending xnn060_val job (7849692) compares the two envs. The same mistake was avoided on ChemAI and missed here. Restored at Paul’s word (2026-10-04, job 7849692 still pending): pip install --no-deps "xnns @ git+https://github.com/molssi-ai/xnn@a8958df6..." into seamm-lammps-xnndev; its pip freeze is now identical to the development session’s envbuild/freeze_dev.txt and pip check is clean. The openssl 3.6.5 patch and the two markers were left.

Test jobs#

  • ChemAI job 5165, queue ChemAI (local SLURM): the MOPAC table flowchart; same table as every earlier run; the table is in its seamm.db.

  • ChemAI job 5166, queue tinkercliffs_debug (SLURM job 7850707 on TinkerCliffs, run in the new environment): same table.

Left for Paul: xnn-step on ChemAI (2026.9.28), to be updated deliberately without applying seamm-xnn.yml to seamm-lammps; the pyxtal-step installer’s pkg_resources import.

MBE on ChemAI#

Paul’s ask, same day: install mbe-step and seamm-mbe on ChemAI. seamm-manager install refused both (“not a SEAMM package”): they are not yet in the nightly package list. They were added to the current venv (venvs/2026-10-04T06-18-41) in place with uv pip install --python ~/SEAMM/venv/bin/python mbe-step==2026.10.4 seamm-mbe==2026.10.3.1, with nothing queued that used the venv. Their dependencies were already there, so nothing else changed; uv pip check is clean and MBE is registered under org.molssi.seamm. mbe-step has no installer (no code environment). No service restart: each job imports the plug-ins afresh. Until the two are in the package list, a later update --all keeps them but environment recreate would drop them.

Phase 5 rollout (2026-10-04, afternoon)#

Paul’s explicit ask (relayed by the design session, confirmed by Paul): ChemAI and TinkerCliffs together, by explicit package list only – never --all, no xnn-step/lammps-step, no code installers (none of the eight packages has one).

Package

Before

After

molsystem

2026.10.3

2026.10.4

seamm

2026.10.3

2026.10.4

seamm-scheduler

2026.10.3

2026.10.4

seamm-exec

2026.10.4

2026.10.4.1

loop-step

2026.10.3

2026.10.4

read-structure-step

2026.10.4

2026.10.4.1

forcefield-step

2026.9.27

2026.10.4

seamm-jobserver

2026.10.3

2026.10.4

Same on both sites. seamm-manager update --latest molsystem seamm seamm-exec loop-step read-structure-step forcefield-step seamm-jobserver; seamm-scheduler is not in the package list (“not a SEAMM package”) and came in through seamm-exec’s requirement. uv pip check clean on both.

  • TinkerCliffs (/projects/seamm/SEAMM, no services): none of the 50-odd running jobs used the venv (ORCA/VASP farm scripts under /projects/seamm/psaxe, checked by each job’s command). New venvs/2026-10-04T15-23-39; rollback venvs/2026-10-04T06-12-15.

  • ChemAI: nothing in SLURM for seamm, nothing running or submitted in the datastore (2482 jobs; the 112 stale “started” rows as before). JobServer and web UI stopped, updated: new venvs/2026-10-04T15-37-21; rollback venvs/2026-10-04T06-18-41. mbe-step 2026.10.4, seamm-mbe 2026.10.3.1 and the held xnn-step 2026.9.28 carried over unchanged. The update reported starting the JobServer but it was not running (seamm_manager#26 again); started it, then the web UI; all 16 queues listed.

Test jobs (job 5165’s flowchart, three molecules from SMILES, MOPAC):

  • ChemAI job 5167, queue ChemAI: finished.

  • ChemAI job 5168, queue tinkercliffs_debug (run in TinkerCliffs’ new venv): finished, identical to 5167.

Both differ from this morning’s 5165/5166 (e.g. -11.3502818 vs -11.3070296 kcal/mol) because of molsystem 2026.10.4’s fixed RDKit seed: the SMILES give other starting conformers. Shown, not assumed: the same flowchart run locally with molsystem 2026.10.3 reproduces 5165/5166 and with 2026.10.4 reproduces 5167/5168, to every printed digit.

Since both sites moved together, no max_resubmits change was needed on ChemAI’s ARC queues.

MBE on TinkerCliffs (2026-10-04, Paul’s ask)#

So that MBE runs use the main install instead of the MBE session’s private venv: uv pip install mbe-step==2026.10.4 seamm-mbe==2026.10.3.1 into /projects/seamm/SEAMM/venv (venvs/2026-10-04T15-23-39), in place as on ChemAI this morning (seamm-manager install refuses packages not in the package list). A dry run first showed only these two; nothing used the venv. uv pip check clean; MBE registered under org.molssi.seamm and listed by seamm-flowchart steps. vasp.ini there (changed by the MBE session with Paul’s OK; backup vasp.ini.bak-2026-10-04) names dftd4 = /projects/seamm/conda-envs/dftd4/bin/dftd4; neither package has an installer, so nothing rewrote it. Until mbe-step and seamm-mbe are in the package list (seamm_packaging PR #3), environment recreate would drop them.